Memo 03 — Ethics as a Source of Norms

Series: The Landscape of Normative Systems (LNS) Memo number: 03 of 12 Primary JD questions: Q1 (sources of normative authority), Q2 (six-part analysis per source), Q8 (what must stay above the kernel). Secondary: Q3 (the norm-generated-by-optimization and norm-generated-by-universalizable-rule behavior classes), Q7 (relation of ethics to law, religion, culture). Status of D0: Immutable input. This memo neither modifies, criticizes, nor defends D0. It maps structure above the kernel.

Scope note. This memo treats morality and ethics strictly as norm-generation mechanisms. It works in the register of moral philosophy, metaethics, and the sociology of moral codes — not as advocacy for any moral position. Its object is structure: how a class of normative systems takes a value, a rule, a character-ideal, or a hypothetical agreement as its root and produces obligations, permissions, and prohibitions that subjects treat as binding independent of external coercion. The memo distinguishes three layers that are routinely conflated in ordinary speech — descriptive morality (what a group in fact holds), normative ethics (accounts of what ought to be held), and metaethics (what moral claims are, and whether they can be true) — because the three layers have sharply different computational signatures. It analyzes the major normative frameworks as norm-generating functions: consequentialism/utilitarianism, deontology, virtue ethics, contractualism, and care ethics. It then applies the six-part analysis (JD Q2) to two concrete sources: morality (as a diffuse social source) and professional/ethical codes (as a formalized institutional source). Its central engineering payload is an argument, built for Memos 08, 11, and 12, that substantive ethical content (e.g., "justice," "human rights," a particular theory of the good) must live above the kernel and that the is/ought gap and the naturalistic fallacy mark a computational boundary the kernel must not cross. Per §0.4 of the conventions, every substantive claim is tagged; when uncertain we tag down. Per M4, we treat moral vocabulary as presentation and look for the computational object beneath it.

3.0 Orientation: why ethics is a distinct — and dangerous — behavior class

Most sources surveyed in this series derive a norm's authority from a procedure that is, at least in principle, observable and repeatable: a vote (Memo 01), a precedent chain (Memo 02), a revelation fixed in a source-object (Memo 04), a contract or a market equilibrium (Memo 05), a convergent convention (Memo 06). Ethics is the paradigm case of a source that claims authority without any such procedure being constitutive of it. When a moral agent asserts "torturing the innocent for amusement is wrong," the assertion does not present itself as valid because someone voted, revealed, contracted, or converged on it. It presents itself as binding regardless of what anyone decided. INFERENCE That self-presentation — authority claimed as procedure-independent and often as agent-independent — is what qualifies ethics as its own behavior class rather than a subspecies of authority, convention, or revelation. INFERENCE

This makes ethics uniquely important and uniquely dangerous for a computable governance project, for one reason: ethics is the source most tempting to build directly into the kernel, and the source that most catastrophically must not be. Legislation is obviously particular; nobody proposes hard-coding a specific tax statute into a content-neutral core. But "justice," "human dignity," and "harm minimization" feel universal, and precisely that feeling is the failure mode this memo is written to prevent. The bulk of the memo's analytic weight (§3.6, §3.9) is spent showing that the universality-feeling is not itself evidence of the content-neutrality the kernel requires.

Vocabulary reminder (from §0.6). We distinguish legitimacy (subjects treat the norm as binding independent of coercion), validity (formal membership of a norm in a system's rule-set), and efficacy (actual compliance). Ethics is the extreme case where legitimacy is claimed at maximum strength while validity is often undefined (there is no agreed rule of recognition for moral norms — contrast Memo 02 on Hart) and efficacy is highly variable. Holding these three apart is where the engineering lives.

We also fix a distinction used throughout:

The three layers answer different questions and carry different computational signatures. Conflating them is the single most common error in "ethical AI" and "value alignment" discourse INFERENCE, and the conventions' M4 ("language is presentation") is the discipline that keeps them apart here.

3.1 Established consensus

Claims here are broadly agreed as descriptions of the intellectual terrain within moral philosophy and the empirical study of morality. Agreement is about what the positions are, what the classic arguments say, and what history recorded — not about which moral theory is correct. No consensus on the latter exists, and its absence is itself one of the load-bearing facts (§3.1.5).

3.1.1 Ethics is, among other things, a normative system

FACT Every human society that has been studied carries a body of moral prescriptions — obligations, permissions, prohibitions — governing killing, harm, honesty, reciprocity, fairness, property, sexuality, purity, and in-group loyalty. Whatever else morality is, it demonstrably produces, maintains, and enforces ought-statements, which is this series' definition of a normative system (§0.6).

INFERENCE Because morality issues ought-statements with subjects, conditions, and modalities, the Hohfeldian primitives (§0.6) apply: moral duties with correlative claim-rights ("you owe me the truth"), moral privileges ("I may refuse"), and — more weakly — moral powers (promising creates an obligation; consent removes one). The fit is looser than for law because moral systems typically lack a determinate authority to allocate the relations, but the eight relations describe moral norms with only modest residue. INFERENCE — provisional, per §0.6.

3.1.2 The descriptive / normative / metaethical tripartition is standard

FACT Moral philosophy standardly separates three inquiries:

1. Descriptive ethics — the empirical study of the moral beliefs and practices a group actually holds. It is a branch of the social sciences; its claims are ordinary empirical claims and are falsifiable in the M1 sense.
2. Normative ethics — the study of what one ought to do and which character or acts are good/right; it proposes and defends first-order standards (the frameworks in §3.3).
3. Metaethics — the study of the status of moral claims: their meaning, their metaphysics (are there moral facts?), and their epistemology (how, if at all, could we know them?).

FACT These are distinct questions. A complete answer to one under-determines the others: two people can share a descriptive account of what their society believes, disagree about what it ought to believe (normative), and disagree again about whether "ought" claims can even be true (metaethics).

INFERENCE The tripartition maps onto the D0-relevant layering as follows, and this mapping is a primary result of the memo:

INFERENCE Only the descriptive layer has an uncontested truth-value. The normative layer is computable only after someone supplies the values it optimizes over or the rules it applies — the kernel cannot manufacture those (this is the whole of §3.6). The metaethical layer determines whether the normative layer's outputs are "true" or merely "valid-relative-to-a-theory"; that question is open (§3.4, §3.7).

3.1.3 The is/ought gap (Hume) is a recognized structural feature

FACT The observation associated with David Hume (A Treatise of Human Nature, 1739–40) is that arguments commonly slide from premises joined by "is" and "is not" to a conclusion joined by "ought" or "ought not," and that this transition is not explained. The passage is standardly read as flagging that a purely descriptive premise set does not, by the ordinary rules of inference, entail a prescriptive conclusion. FACT that the passage exists and is read this way; the exact strength of Hume's own intended claim is debated, so we tag the interpretation down. INFERENCE

INFERENCE Stated formally, the gap is a claim about the closure of logical consequence: if every premise is descriptive (contains no ought), then no non-trivial prescriptive conclusion follows, because a valid deductive inference cannot introduce vocabulary essential to the conclusion that is absent from and undefined by the premises. This is often called Hume's Law. Its formal status is contested (see §3.5.1) but the default engineering reading is: you cannot derive an obligation from facts alone; an obligation-yielding premise (a value, a rule, a goal) must already be present. INFERENCE

INFERENCE — this memo's central boundary claim, stated here, defended in §3.6 Hume's gap is, for our purposes, a computational boundary: it marks the input port at which a value must be injected from above the kernel. The kernel can compute consequences of an ought; it cannot compute an ought from an is. Any architecture that appears to derive substantive obligations from purely factual inputs has smuggled a value in through an unlabeled port.

3.1.4 The naturalistic fallacy and the open-question argument are recognized

FACT The term "naturalistic fallacy" is associated with G. E. Moore (Principia Ethica, 1903). Moore's target is the identification of the property good with any natural property (e.g., pleasant, desired, conducive to survival). His open-question argument runs: for any proposed natural definition "good = N," the question "but is N good?" remains intelligibly open, whereas "is good good?" is trivial; therefore good is not identical to N. FACT that this is Moore's argument.

INFERENCE Two cautions the series must carry: - Hume's gap (is→ought, a claim about inference) and Moore's naturalistic fallacy (a claim about the definition/identity of moral properties) are distinct, though often run together. One is about entailment; the other is about property identity. FACT that they are distinct. - The open-question argument is itself contested (§3.5.2); it does not prove non-naturalism. We use it only in its weak, defensible form: the reduction of a moral term to a specific natural quantity is never self-evidently correct and always invites the further normative question. INFERENCE

INFERENCE Engineering translation: any attempt to define a governance target as a measurable natural quantity ("welfare = summed reported satisfaction," "harm = count of X") inherits the open question "but is maximizing that quantity good?" — which is a value question the kernel cannot close. This is the naturalistic fallacy re-encountered as a specification problem (§3.9.3, and the Goodhart connection in §3.3.1).

3.1.5 Persistent, reasonable disagreement is an established datum

FACT There is no convergence, after millennia of inquiry, on a single correct normative theory, nor on a single correct metaethics. Trained, informed, good-faith philosophers persistently disagree. This is not itself a moral claim; it is a sociological observation about the field.

INFERENCE The persistence of disagreement is evidence under M8 (theories are evidence, not authority): it constrains the kernel design directly. A kernel that had to select one normative theory to function would inherit a dispute that the most capable inquirers have not resolved. Under M6 (discovery, not invention) we may not resolve it by fiat. This drives the exclusion argument of §3.6 and connects to the political fact of pluralism handled in Memo 01 and the Rawlsian "burdens of judgment" noted in §3.3.4. INFERENCE

3.1.6 The frameworks agree far more at the level of cases than of theory

FACT Across cultures and across normative theories, a core of prohibitions recurs: against unprovoked killing of in-group members, gratuitous cruelty, betrayal of trust, theft, and deception, alongside positive expectations of reciprocity and care for dependents. Comparative and empirical work (in the anthropology of morality and moral psychology) documents this recurrence. FACT — as an empirical regularity; its explanation is contested.

INFERENCE The theories diverge most at the justification layer and at the hard cases (trade-offs, rights vs. aggregate welfare, partiality vs. impartiality), and converge most at the paradigm-case layer. This is structurally important: it suggests the computable invariant, if any, is more likely found in the shape of moral reasoning (a source feeding a rule/valuation feeding a verdict, with conflict-resolution) than in the content of any verdict. This is the same "clean shape / non-neutral content" split identified for religion in Memo 04, and it is developed in §3.8. HYPOTHESIS

3.2 Six-part source analysis (JD Q2)

The conventions require, for each source, the analysis: origin, legitimacy, enforcement mechanism, mutation mechanism, hierarchy, conflicts, computational implications. Ethics is not a single source; it is a family. We analyze two concrete members whose computational signatures differ sharply: (A) morality as a diffuse, non-institutional social source, and (B) professional/ethical codes as a formalized institutional source. The normative theories of §3.3 are analyzed there as the norm-generating functions these sources may run.

3.2.A Source: MORALITY (diffuse social source)

Origin. INFERENCE Morality-as-practiced has no single origin. The scholarly candidates (treated under M8 as competing, possibly all-partial): (i) descriptive-cultural transmission (morality as inherited convention — links to Memo 06); (ii) an evolved cooperation substrate (moral emotions and reciprocity dispositions selected because they stabilize cooperation) HYPOTHESIS — an empirical claim, falsifiable, not settled; (iii) rational reflection converging on constraints (the position associated with Kantian and contractualist traditions) INFERENCE; (iv) response to mind-independent moral facts (moral realism, §3.4.1) [contested]. For source-analysis purposes the key FACT is that morality presents itself to its subjects as found, not made — even where a social scientist would describe it as constructed.

Legitimacy. INFERENCE Moral norms achieve the strongest legitimacy signature in the taxonomy: subjects treat them as binding independent of coercion, of self-interest, and often of authority ("even if the law permitted it, it would still be wrong"). This procedure-independence (§3.0) is morality's defining legitimacy feature. Crucially, legitimacy here is decoupled from validity: there is no agreed rule of recognition (contrast Hart, Memo 02) that certifies a moral norm as "valid," so validity is largely undefined and legitimacy does all the work. Efficacy is variable and enforced softly (below).

Enforcement mechanism. FACT /INFERENCE Morality's enforcement stack is predominantly decentralized and informal:

The first-party layer (conscience) is what distinguishes moral from merely conventional enforcement INFERENCE; a norm enforced only externally is closer to convention (Memo 06) than to morality. There is no central detector and no authorized adjudicator, which is the root of morality's computational difficulty (below).

Mutation mechanism. FACT Moral codes change, sometimes fast (documented shifts over the last two centuries in the moral status of slavery, of punishment, of the moral standing of previously excluded groups). Mechanisms: gradual attitude drift; moral entrepreneurship and argument; generational turnover; expansion or contraction of the circle of moral concern; and feedback from law and religion (Memos 02, 04). INFERENCE There is no amendment procedure: mutation is emergent, not enacted. This is the opposite of legislation (Memo 01) and the opposite of a revelation's fixed root (Memo 04); it most resembles convention drift (Memo 06).

Hierarchy. INFERENCE Internally, morality lacks a formal hierarchy of norms; instead it exhibits priority intuitions (some duties felt as near-absolute — e.g., against gratuitous cruelty — others as defeasible) and framework-level disputes about what grounds priority (rights-trump-welfare vs. welfare-aggregation, §3.3). Externally, the hierarchy between morality and law is itself a jurisprudential battleground (natural law vs. positivism, Memo 02): does an immoral law lack validity, or is it valid-but-to-be-resisted? OPEN — this is not settled and the kernel must not pre-decide it, §3.6.

Conflicts. INFERENCE Morality generates conflicts of three kinds: (1) intra-personal (competing duties in one agent — the classic tragic dilemma); (2) inter-personal / inter-group (differing moral codes across communities — moral disagreement, §3.1.5); (3) inter-source (moral norm vs. legal, religious, professional, or market norm — e.g., civil disobedience, conscientious objection). Morality provides no authoritative conflict resolver; resolution is by persuasion, power, drift, or migration into a harder source. INFERENCE

Computational implications. INFERENCE - Morality supplies content and legitimacy pressure, not a decision procedure. It cannot be executed as-is: no detector, no adjudicator, no amendment port. - Its truth-value status is unsettled (§3.4), so a kernel cannot assume moral claims are truth-apt propositions it may evaluate. - What is computable about morality is second-order: representing that a population holds norm X (descriptive layer, §3.1.2), tracking priority and conflict relations, and recording the mapping from a chosen framework to verdicts. The first-order content ("X is wrong") is an input, not an output. - Design consequence: morality enters the system as a typed value input with provenance, never as a kernel axiom. (Elaborated §3.6, §3.9.)

3.2.B Source: PROFESSIONAL / ETHICAL CODES (formalized institutional source)

Examples in the register of comparative professional regulation: medical codes (the tradition associated with the Hippocratic oath and modern successors), legal professional conduct rules, engineering codes of ethics, accounting and auditing standards, research-ethics frameworks, and corporate codes of conduct. FACT that such codes exist and are enforced by professional bodies.

Origin. FACT Codes are authored by an institution — a professional association, licensing board, standards body, or firm — typically claiming to codify a pre-existing ethical consensus of the profession plus the public interest. INFERENCE Structurally this is a hybrid: it borrows content from morality (the diffuse source, 3.2.A) but acquires validity from an institutional enactment procedure (like legislation, Memo 01/07). This makes codes the cleanest bridge in the ethics family between the un-executable diffuse source and an executable rule-set.

Legitimacy. INFERENCE Dual-rooted: (i) moral legitimacy inherited from the underlying ethical content ("this is what a good doctor owes patients"), and (ii) institutional legitimacy from the authorized body and, often, from state delegation (licensure). The two can diverge: a code may be institutionally valid yet morally contested (e.g., a code lagging behind shifted moral consensus), or morally compelling yet institutionally unadopted. Holding legitimacy ≠ validity apart is essential here.

Enforcement mechanism. FACT Unlike diffuse morality, professional codes have a concrete enforcement stack: complaint procedures, disciplinary tribunals, sanctions (reprimand, fines, suspension, loss of license), and sometimes onward referral to civil or criminal law. This is a genuine detector + adjudicator + consequence pipeline — i.e., an executable enforcement layer, which diffuse morality lacks. FACT /INFERENCE

Mutation mechanism. FACT Codes mutate by enacted amendment: committees revise, boards ratify, versions are published with effective dates. This is much closer to legislation's mutation (Memo 01) than to moral drift. INFERENCE Consequently codes are versionable and diffable — a property of first-order importance for D1 (§3.9).

Hierarchy. FACT /INFERENCE Codes sit in a definable hierarchy: statute and regulation typically override a professional code where they conflict; the code overrides individual firm policy; firm policy overrides personal preference. Internally, many codes rank principles (e.g., patient welfare / autonomy ahead of professional convenience) or at least flag which duties are non-waivable. This gives codes something diffuse morality lacks: partially specified precedence.

Conflicts. INFERENCE Codes generate characteristic conflicts: duty-to-client vs. duty-to-public; confidentiality vs. disclosure; role-obligation vs. personal moral conviction (the conscientious-objection case, again). Some codes include tie-break or escalation rules (report to a named authority); many do not, leaving residual dilemmas. Where a code conflicts with law, the hierarchy above usually resolves it; where it conflicts with personal morality, it does not.

Computational implications. INFERENCE - Professional codes are the most kernel-tractable member of the ethics family because they already have: an enactment/versioning procedure (validity), a determinate authority (adjudicator), an enforcement pipeline, and partial precedence. They look, computationally, more like regulation (Memo 07) than like morality. - They are therefore a good test corpus for D1: if the intermediate representation cannot express a real professional code (its obligations, permissions, prohibitions, precedence, exceptions, and enforcement hooks), it is inadequate. HYPOTHESIS — proposed as an acceptance test. - But their content remains substantive ethics (what patients are owed, what counts as harm) and thus stays above the kernel exactly as diffuse morality does (§3.6). The code's form is executable; the code's values are inputs. This form/content split is the recurring result.

3.2.C Contrast table

INFERENCE The table isolates the payload for D1/D2: the executability of an ethical source scales with how much institutional apparatus (validity rule, adjudicator, enforcement, versioning, precedence) it carries — none of which is ethical content. The content is constant across the row; only the apparatus varies. This is the empirical backbone of the exclusion argument (§3.6).

3.3 Competing theories I: normative frameworks as norm-generating functions

The engineering reframing (per M4): a normative theory is a function that maps some inputs to a verdict (right/wrong, obligatory/permitted/forbidden, or a ranking). The theories differ in what they take as input and what they optimize or check. Presenting them this way exposes exactly where each one requires a value to be supplied from outside — the input port that cannot be kernel-internal (§3.6). We do not adjudicate between them (M6, M8); we characterize their signatures and failure modes.

General form:

$$\text{verdict} = F_{\text{theory}}(\text{act} \mid \text{agent} \mid \text{situation} \mid \text{VALUES})$$

where the $\text{VALUES}$ argument is supplied from above the kernel.

Each subsection gives: the function, its inputs, its output type, its computational tractability, and its standard destructive tests (M5).

3.3.1 Consequentialism / utilitarianism — norm from optimization of an aggregate value

Function. FACT — as a characterization of the family An act (or rule) is right iff it produces the best consequences, where "best" is fixed by a value function V over outcomes and an aggregation A over affected parties: $$\mathrm{right}(a) \equiv a \in \operatorname*{argmax}_{a' \in A}\left[\ \mathrm{Aggregate_over_persons}\big(V(\mathrm{outcome}(a'))\big)\ \right]$$ Classical utilitarianism (the tradition associated with Bentham and Mill) sets V = welfare/utility and A = sum. Variants change V (preference-satisfaction, objective goods) or A (average vs. total; priority-weighting for the worse-off).

Inputs that must be supplied from above the kernel. (1) the value function V (what is good); (2) the aggregation rule A (how goods across persons combine — a distributive value); (3) the scope of moral patients (whose good counts); (4) a prediction model outcome(·). INFERENCE Items (1)–(3) are pure values; item (4) is empirical. Only (4) is on the "is" side of Hume's gap.

Output type. A total (or partial) preference ordering over acts, collapsible to a verdict. This is consequentialism's engineering attraction: it natively yields a ranking, hence an optimization target, hence something an executor can maximize. INFERENCE

Computational tractability. INFERENCE Superficially the most "computable" framework — it is literally an optimization. But three hard limits: - Value non-derivability. V and A are inputs; nothing in the machinery selects them. Choosing V is the naturalistic-fallacy trap (§3.1.4): fixing "good = measurable quantity Q" leaves open "is maximizing Q good?" - Prediction/compute. outcome(·) over open-ended futures is generally intractable/uncomputable in the limit; consequentialism assumes an oracle it does not have. INFERENCE - Goodhart failure. FACT that this is a recognized effect When a measured proxy for V is optimized hard, the proxy decouples from the intended good ("when a measure becomes a target, it ceases to be a good measure"). This is the naturalistic fallacy re-encountered dynamically: the specification gap becomes an exploit under optimization pressure. This connects directly to KPI gaming (Memo 07) and to the specification-gaming risk in Memo 11.

Destructive tests (M5). The standard counterexamples: the demand to punish an innocent when aggregate welfare would rise (rights have no independent standing); the "utility monster"; the repugnant conclusion under total aggregation; the collapse of the act/rule distinction under pressure. These do not refute the family but show its verdicts are hostage to V and A, i.e., to inputs. INFERENCE

3.3.2 Deontology — norm from universalizable rule / duty

Function. FACT — as characterization An act is right iff it conforms to a rule that passes a test of the rule itself, independent of consequences. The paradigm test associated with Kant: an act's maxim is permissible iff it can be universalized without contradiction (the categorical imperative in its universal-law and humanity-as-end formulations). Rule-based deontologies more generally supply a set of duties/constraints and check membership: $$ \begin{aligned} \mathrm{permitted}(a) &\equiv \mathrm{maxim}(a) \in {\, m : \mathrm{universalizable}(m) \land \mathrm{treats_persons_as_ends}(m) \,} \ \mathrm{right}(a) &\equiv a \text{ satisfies all applicable duties } D \text{ (and violates no constraint)} \end{aligned} $$

Inputs from above the kernel. (1) the duty set D or the universalizability test; (2) the correct description of the maxim/act (the individuation problem — under which description is the act tested?); (3) a priority rule among conflicting duties. INFERENCE The universalizability test looks content-neutral and procedural — which is why deontology is the framework most often mistaken for something that could be kernel-internal. §3.6.2 argues that this appearance is misleading: the test's verdicts depend entirely on maxim individuation (2), which is not procedural.

Output type. A deontic classification: obligatory / permitted / forbidden (the O/P/F operators of §0.6) — natively categorical, not a ranking. This makes deontology map cleanly onto constraint checking rather than optimization. INFERENCE

Computational tractability. INFERENCE - Constraint-checking is computationally cleaner than optimization: given the rule set and the act's description, membership is often decidable. - But maxim individuation is the killer: "I am lying to a murderer at the door to save a life" universalizes differently than "I am lying." Nothing in the test fixes the description; the description encodes a value judgment. - Duty conflicts (the tradition associated with Ross's prima facie duties) require a priority rule the theory does not fully supply → residual dilemmas. - Deontic logic (§0.6) formalizes the O/P/F output but inherits the known paradoxes (Ross, Chisholm contrary-to-duty, gentle murderer), treated as destructive tests in Memo 10.

Destructive tests (M5). The murderer-at-the-door (absolutism about lying yields monstrous results); irresolvable duty conflicts; the individuation regress. Again these show the verdicts depend on inputs (D, the individuation convention) rather than being generable by the test alone. INFERENCE

3.3.3 Virtue ethics — norm from character / telos

Function. FACT — as characterization The primary bearer of evaluation is the agent's character (virtues/vices), not the act or its outcome. Rightness is derivative: an act is right iff it is what a person of practical wisdom (phronesis) with the relevant virtues would characteristically do, given a conception of human flourishing (eudaimonia / telos): $$ \begin{aligned} \mathrm{right}(a \mid \text{situation}) &\equiv a = \mathrm{what_a_virtuous_agent_would_do}(\text{situation}, \text{TELOS}) \ \mathrm{good_agent} &\equiv \text{agent whose stable dispositions realize the virtues} \end{aligned} $$ Associated primarily with the Aristotelian tradition and its modern revivals.

Inputs from above the kernel. (1) the list of virtues; (2) the conception of the telos/flourishing that fixes which traits are virtues; (3) phronesis — context-sensitive practical judgment, explicitly held to be non-algorithmic by the tradition itself. FACT that the tradition characterizes phronesis as not reducible to rules.

Output type. INFERENCE Not a verdict function of the same type as the others. Virtue ethics is agent-centric and non-codifiable by design: it resists reduction to a decision procedure and instead outputs an evaluation of character plus context-dependent guidance. This is not a defect in its own terms — it is a thesis that the moral is not fully rule-capturable.

Computational tractability. INFERENCE - Virtue ethics is the framework most resistant to computation, and deliberately so: its central claim (uncodifiability of practical wisdom) is, if true, itself a limit result about the ethics family — a candidate falsifier for "morality is fully computable" (relevant to Memo 11's attack). HYPOTHESIS — the uncodifiability claim is contested, not established. - What it does contribute computationally: the observation that a governance system may need to evaluate dispositions/agents over time (trustworthiness, track record), not only discrete acts. This maps to reputation and role-based design (Memos 06, 07). INFERENCE

Destructive tests (M5). The "situationist" empirical challenge (whether stable cross-situational character traits exist as the theory assumes) is an empirical falsifier on the descriptive layer; the "self-effacing"/action- guidance objection (it tells you to be good without telling you what to do) is a structural one. Both are live. INFERENCE

3.3.4 Contractualism — norm from hypothetical agreement

Function. FACT — as characterization A norm is justified iff it would be agreed to by suitably situated parties under specified conditions: $$\mathrm{justified}(\text{norm}) \equiv \text{norm} \in {\, n : \text{all parties } P \text{ in condition } C \text{ would accept } n \,}$$ Two prominent variants (attributed by name only as safe): the Rawlsian device of choosing principles behind a veil of ignorance in an original position (so no one tailors principles to their own position); and the Scanlonian criterion that a principle is wrong if it could be reasonably rejected by someone. Also the descriptively-flavored contractarian tradition (Hobbesian / bargaining-from-self-interest) which is closer to game theory (Memo 05).

Inputs from above the kernel. (1) the specification of the parties and the choice condition C (the veil's thickness, what parties know/want, what counts as "reasonable"); (2) the account of parties' interests; (3) the standard of "reasonable rejection." INFERENCE These specifications are substantive value choices — different C's yield different principles. Contractualism does not remove the value input; it relocates it into the design of the choice situation. This is the same move deontology makes with maxim individuation and is critical for §3.6.

The burdens of judgment. INFERENCE The Rawlsian observation that free reason, under normal conditions, predictably yields reasonable pluralism about the good — the "burdens of judgment" — is itself an argument that no single comprehensive doctrine can be the shared public basis. This is a philosophical mirror of §3.1.5 and it points toward keeping comprehensive ethical content out of any shared core (the "political, not metaphysical" strategy), which is the intra-philosophical analogue of the exclusion argument (§3.6). INFERENCE

Output type. A justified set of principles/norms (not a per-act ranking). Contractualism operates at the level of rules/institutions, which makes it the framework most naturally aligned with constitutional and procedural design (Memo 01, Memo 12). INFERENCE

Computational tractability. INFERENCE - The agreement test can be modeled game-theoretically (Memo 05) and is, in simple settings, tractable; the veil-of-ignorance move is even implementable as a symmetry/impartiality constraint on rule selection. - But the output is only as neutral as C, and specifying C is a value act. The tractable core is the impartiality/symmetry structure; the substantive principles are inputs.

Destructive tests (M5). Indeterminacy (many principle sets survive the test); the "circle" objection (the choice conditions are rigged to yield the theorist's preferred principles); exclusion of parties who cannot bargain (future generations, non-human animals, the severely disabled) — which is precisely the gap care ethics targets. INFERENCE

3.3.5 Care ethics — norm from relationship and responsiveness

Function. FACT — as characterization Moral salience originates in concrete relationships and the responsibilities of care they generate, rather than in impartial rules, aggregate welfare, or hypothetical contracts. The right response is attentive, responsive care for particular others in actual relations of dependence: $$\mathrm{obligation}(\text{agent} \to \text{other}) \sim f\big(\mathrm{relationship}(\text{agent}, \text{other}),\ \mathrm{need}(\text{other}),\ \mathrm{capacity}(\text{agent})\big)$$ Associated with a tradition emphasizing relational, contextual, partiality- respecting moral reasoning (developed substantially in late-20th-century moral philosophy).

Inputs from above the kernel. (1) the map of relationships and dependencies; (2) the account of needs; (3) a conception of appropriate responsiveness. INFERENCE

Output type. INFERENCE Context- and relationship-indexed obligations; like virtue ethics, resistant to impartial codification and explicitly critical of the impartiality assumption shared by consequentialism, deontology, and contractualism.

Computational tractability. INFERENCE - Care ethics is important to this memo less as a computable function than as a falsifier of a hidden assumption: that morality is impartial and agent-neutral. If some genuine obligations are essentially partial (relationship-indexed), then any kernel that assumes agent-neutral, position-independent norms has excluded a real class of norms by construction. HYPOTHESIS This is a design warning, not a settled result. - Constructively, it maps to the fact that governance systems do encode special obligations (fiduciary duties, guardianship, duty of care) that are relationship-indexed — so the representation must at least support role-and-relation-indexed obligations, even though their content is supplied from above (Memos 05, 07). INFERENCE

3.3.6 The common form across all five frameworks

INFERENCE — a central structural result of the memo Abstracting over §3.3.1–5:

$$\text{every normative framework} = (\ \mathrm{SELECTOR},\ \mathrm{VALUE\text{-}INPUT},\ \mathrm{CONFLICT\text{-}RULE}\ )$$

Three observations follow:

1. The SELECTOR is (partly) formal; the VALUE-INPUT never is. In every framework the machinery is content-neutral until the value input is supplied, and no framework's machinery generates its own value input. This is Hume's gap appearing five times in five costumes. INFERENCE

2. The frameworks are not merely rival answers; they disagree about the type signature of the moral function (rank vs. classify vs. evaluate- agent vs. justify-rule vs. respond-to-relation). A kernel that picked one type signature would silently exclude the others. INFERENCE

3. The CONFLICT-RULE is under-specified in all five. None supplies a complete, non-arbitrary resolution for genuine dilemmas. This is not a contingent gap to be patched; it recurs everywhere and is a candidate invariant incompleteness of normative content (connects to Memo 10's deontic paradoxes and Memo 11's attack). HYPOTHESIS

The payload for D1/D2: what is potentially computable is the SELECTOR + CONFLICT-RULE machinery as a parameterized shape, with VALUE-INPUT as a typed argument. What is not computable-from-within is the VALUE-INPUT itself. §3.6 turns this into the exclusion rule.

3.4 Competing theories II: metaethics and the truth-value question

Metaethics decides a question the kernel cannot dodge: when a moral norm is handed to the system, is there a fact of the matter that makes it true or false, or is the norm merely an attitude, a command, or a construction? The answer changes what "evaluating" a norm could even mean. We survey the four principal positions (M8: evidence, not authority) and extract, for each, the computational consequence. We do not decide among them; the point is that the kernel must function under uncertainty about which is correct (§3.7).

Two orthogonal axes organize the field: | Axis | Question | Values | |------|----------|--------| | COGNITIVISM | Are moral sentences truth-apt (express beliefs)? | yes = cognitivist ; no = non-cognitivist | | OBJECTIVITY | If truth-apt, are any of them true mind-independently? | yes = realist ; no = anti-realist/constructivist |

3.4.1 Moral realism — there are moral facts

FACT — as position description Moral realism holds that moral sentences are truth-apt (cognitivism) and that some are true in virtue of mind-independent moral facts. Sub-variants: non-naturalist realism (moral facts are sui generis, not reducible to natural facts — the tradition associated with Moore and with later intuitionism) and naturalist realism (moral facts are natural facts of some kind).

Computational consequence. INFERENCE If realism is true, moral norms have determinate truth-values in principle — the kernel could in principle host an evaluation function. But even granting realism: - Non-naturalist realism explicitly denies the facts are natural, hence they are not measurable/observable and give the kernel no accessible oracle. Truth exists but is epistemically out of computational reach. INFERENCE - Naturalist realism must identify which natural facts — reopening the open-question argument (§3.1.4) and the specification problem. INFERENCE - Realism does not deliver agreement (§3.1.5): even if there are moral facts, we lack a decision procedure to read them, so persistent disagreement stands. So realism-being-true would not license building moral content into the kernel; it would at most license an above-kernel evaluator whose verdicts remain contested. INFERENCE

3.4.2 Anti-realism: non-cognitivism and subjectivism

FACT — as position description Anti-realist families: - Non-cognitivism (emotivism; prescriptivism; contemporary expressivism/quasi-realism): moral sentences do not primarily express truth-apt beliefs but attitudes, commands, or endorsements ("Boo/Hurrah"; "Do not do X!"). Quasi-realist variants then try to earn back ordinary truth-talk without positing moral facts. - Subjectivism / relativism: moral sentences are truth-apt but made true by the attitudes of an individual or culture — true relative to a standard, not mind-independently.

Computational consequence. INFERENCE If non-cognitivism is right, then "evaluate whether norm N is true" is a category error: there is no proposition to evaluate, only an attitude to represent. The computable object is then the attitude/endorsement and its holder, i.e., the descriptive layer (§3.1.2) — "agent/group G endorses N" — which does have a truth-value. This is a clean result: on anti-realism, the only truth-apt moral content is descriptive-provenance data, exactly the form §3.2 assigns to morality as a kernel input. INFERENCE Under relativism, the kernel must additionally carry the index (true-for-whom), i.e., norms are parameterized by community — a representational requirement, not a kernel value.

3.4.3 Constructivism — truth as output of a procedure

FACT — as position description Constructivism holds that moral truths are not found (as realism says) nor mere attitudes (as non-cognitivism says) but constructed as the output of a privileged procedure of practical reasoning (the Kantian and Rawlsian constructivist traditions being the standard references). What is right just is what the correct procedure yields.

Computational consequence. INFERENCE Constructivism is the metaethics most congenial to a computational reading, because it locates moral validity in a procedure — and procedures are the kernel's native object. This is significant for Memo 12: it suggests the kernel/above split may line up with the philosophy's own procedure (constructible) vs. inputs-to-the-procedure (given) distinction. But the caution is exact and repeated: the procedure itself embeds substantive commitments (what counts as a reason, who the agents are, what rationality requires). Constructivism relocates the value into the procedure's design; it does not eliminate it. So it licenses a content-neutral procedural kernel with value-laden procedure parameters supplied from above — which is precisely the architecture §3.6 argues for, and precisely why constructivism must not be mistaken for a license to embed a specific construction procedure in the kernel. INFERENCE

3.4.4 Error theory — moral claims are systematically false

FACT — as position description The error theory (associated with J. L. Mackie) is cognitivist but anti-realist in verdict: moral sentences purport to describe objective prescriptive facts, but no such facts exist, so all positive moral claims are uniformly false (as arithmetic would be if there were no numbers). Supporting arguments include the "argument from queerness" (objective prescriptivity would be metaphysically and epistemically anomalous) and the "argument from relativity" (observed disagreement is better explained by projection than by tracking facts).

Computational consequence. INFERENCE If error theory is right, a kernel that evaluated moral truth would return "false" for everything — useless. The only workable stance is again to treat moral norms not as truth-claims to evaluate but as adopted standards to operate under (a fictionalist or practical stance: "we proceed as if"). This lands, once more, on: the kernel processes norms as posited inputs with provenance, not as propositions it verifies. INFERENCE

3.4.5 The convergent computational lesson across metaethics

INFERENCE — a key result The four positions disagree profoundly about the nature of moral claims, yet they converge on a single instruction for the kernel:

CONVERGENCE: on NO defensible metaethics may the kernel treat a substantive moral norm as a self-standing proposition it can verify true/false internally.

INFERENCE This is the strongest form of the exclusion argument's premise: it does not depend on choosing a metaethics. Whatever the correct metaethics turns out to be, the kernel's correct behavior is the same — accept moral norms as typed, provenance-tagged inputs; never as internally-verifiable axioms. This robustness-under-metaethical-uncertainty is exactly what an immutable kernel needs, because the kernel cannot be re-decided when the metaethics debate is (if ever) resolved. This directly serves M6 (an outcome we did not force) and feeds §3.6 and §3.9.

3.5 The is/ought gap and the naturalistic fallacy as computational boundaries

Sections 3.1.3–3.1.4 introduced Hume's is/ought gap and the open-question argument (the position associated with G. E. Moore) as recognized structural features of moral discourse. Here we restate them as computational boundary conditions on any governance compiler — the form in which they matter to D1/D2.

3.5.1 The is/ought gap as a type barrier

INFERENCE The is/ought gap has a clean computational reading: there is no total, truth-preserving function from purely descriptive propositions to normative propositions. Descriptive inputs are of type Fact (world-states, measurements, predictions); norms are of type Norm (obligation / permission / prohibition attached to a subject and condition). No inference rule internal to the kernel can manufacture a value of type Norm from values of type Fact alone. A premise of type Norm must already be present.

forbidden (no such total truth-preserving rule):

$$\frac{\text{Fact},\ \text{Fact},\ \ldots,\ \text{Fact}}{\text{Norm}} \qquad (??\ \text{— cannot be discharged by the kernel})$$

permitted (a norm premise is supplied from above the kernel):

$$\frac{\mathrm{Norm(policy)} \qquad \mathrm{Fact(measurement)}}{\mathrm{Norm(applied\ conclusion)}}$$

INFERENCE This is not a limitation to be engineered away; it is the precise sense in which the kernel must be content-neutral. The kernel can compute Fact → Fact (prediction), and (Norm ∧ Fact) → Norm (application / subsumption), but never Fact → Norm (justification). The ought-premises enter only as declared inputs carrying provenance from a source above the kernel (Memos 01–08). FACT This matches the standard logical statement that a valid deductive argument with a normative conclusion requires at least one normative premise.

OPEN Whether every real governance rule cleanly decomposes into Norm-premise + Fact is not established. Practical rules often smuggle evaluative terms into ostensibly descriptive predicates ("reasonable", "excessive", "safe"). Detecting and quarantining such thick predicates is an open representational problem handed to D1 (see §3.6, §3.8).

3.5.2 The naturalistic fallacy as a prohibition on definitional shortcuts

INFERENCE The open-question argument tells the compiler architect something narrower but sharp: do not hard-code a definition of a moral predicate (e.g., "good", "just", "fair") as though it were a settled equivalence. For any proposed naturalistic definition good(x) ≡ φ(x) (φ some descriptive property — pleasure, preference-satisfaction, evolutionary fitness), the question "but is everything that is φ actually good?" remains substantively open. Encoding such an equivalence in the kernel would bake one contested metaethical theory into a component that is supposed to outlive all such debates — a direct violation of M6.

HYPOTHESIS The corresponding design rule: thick evaluative predicates are represented in D1 as named, provenance-tagged parameters bound above the kernel, never as kernel-level definitions. Falsifiable by exhibiting a moral predicate whose extension is genuinely non-contested across all live metaethical positions (none is currently known).

3.6 JD Q8: why substantive ethical content must stay ABOVE the kernel

This section states the exclusion argument the JD requests (Q8) for the ethical concepts — justice, human rights, fairness, dignity, the good — and connects it to the convergence result of §3.4.5. The argument has three independent legs; any one suffices, and they do not depend on each other.

3.6.1 Leg 1 — Non-neutrality (the content argument)

INFERENCE Substantive ethical predicates are essentially contested concepts (the notion associated with W. B. Gallie): competent, sincere reasoners converge on the word while systematically diverging on its extension. "Justice" names distributive equality for one tradition, desert for another, and Rawlsian fair procedure for a third (see Memo 11 on Rawls/Nozick). FACT There is no agreed algorithm mapping cases to a "just / unjust" verdict across these traditions.

INFERENCE If the kernel contained a definition of justice, it would thereby select one contested tradition and demote the others by fiat. That is precisely the ideological commitment the kernel is required to lack. Therefore justice, human rights, fairness, dignity, and the good are inputs to be supplied and defended above the kernel — never kernel primitives.

3.6.2 Leg 2 — Non-verifiability (the truth-value argument)

INFERENCE From §3.4.5: on no defensible metaethics can the kernel treat a substantive moral norm as a proposition whose truth it can internally verify. Realism makes moral facts real but epistemically inaccessible to a mechanical oracle; non-cognitivism denies they are truth-apt at all; constructivism locates their truth in a procedure that itself sits above the kernel; relativism makes truth index-relative. A kernel primitive would have to presume a verification procedure that no metaethics grants it. Hence such content cannot be a kernel axiom.

3.6.3 Leg 3 — Non-amendability under an immutable kernel (the mutation argument)

INFERENCE Ethical content evolves (chattel slavery once broadly licensed is now near-universally prohibited; see Memo 09 on mutation). The kernel is frozen (D0). Anything placed inside a frozen artifact cannot track a moving target. Therefore any content that legitimately mutates — which includes all substantive ethical content — must live in the mutable layer above the kernel, not inside D0. Placing it in the kernel would either freeze a moral error permanently or force illicit kernel amendment.

3.6.4 What the kernel MAY carry (the positive residue)

INFERENCE The exclusion is of substantive content, not of structure. The convergence of §3.4.5 leaves a content-neutral residue the kernel may legitimately carry, because every ethical framework in §3.3 shares it:

INFERENCE Every framework in §3.3 was shown (§3.3.6) to have the shape (subject, situation-features) → deontic verdict. That shape is content-neutral and is a legitimate candidate for the kernel/interface (developed in Memo 10 and finalized in Memo 12). The binding of the free parameters — which value, which duties, which objective — is the ideological act that stays above the kernel. This is the ethics-specific form of the series' central boundary claim.

3.7 Open questions

• OPEN The thick-predicate decomposition problem. Can every operative governance rule be rewritten as Norm-premise + descriptive-Fact with all evaluative content isolated into named parameters, or do some rules resist decomposition (§3.5.1)? Unresolved; critical for D1's type system.
• OPEN Aggregation of moral inputs. When multiple legitimate ethical sources supply conflicting norms, is there a content-neutral combination rule, or must every combination rule itself smuggle in a substantive value (a weighting)? Connects to Arrow-type limits (Memo 01 §2, Memo 10).
• OPEN Defeasibility. Moral norms are characteristically defeasible (prima facie duties, the notion associated with W. D. Ross): "keep promises, unless…" with an open-ended exception set. Whether defeasible norms admit a finite, sound computational representation is unresolved and recurs in Memo 10.
• OPEN Supererogation and the deontic gap. Acts beyond duty (praiseworthy but not obligatory) sit outside the O/P/F partition. Standard deontic logic represents them awkwardly. Whether the kernel's modality set must be extended is open (feeds Memo 10's stress test of the Hohfeld/deontic vocabulary).
• OPEN Moral uncertainty. How should a governance system act when the correct first-order moral theory is itself uncertain? "Maximize expected choiceworthiness" is one proposal, but it presupposes intertheoretic comparisons that may be meaningless. Unresolved; relevant to any D2 that must act under normative pluralism.

3.8 Research opportunities

• INFERENCE A provenance-typed norm object. D1 should represent each norm with an explicit source tag (which ethical/legal/political source authored it) and an explicit modality; §3.6.4's shape gives the schema. This lets the system carry conflicting moral inputs without collapsing them into a single contested "truth". Directly reusable by Memos 02, 04, 05.
• INFERENCE A thick-predicate registry. Build the (initially manual) catalogue of evaluative predicates that must be bound above the kernel ("reasonable", "fair", "excessive", "cruel"). Serves §3.5.1 / §3.7 and is a concrete six-month D1 task.
• INFERENCE A framework-as-plugin experiment. Because §3.3.6 shows all frameworks share the verdict-function shape, D1 could implement consequentialist, deontological, and contractualist evaluators as interchangeable plugins over the same case representation, and measure where they diverge. This operationalizes M3/M5 (hostile counterexamples appear exactly at the divergences).
• HYPOTHESIS Defeasibility via priority orders. Represent prima facie duties as a defeasible-logic system with an explicit, externally-supplied priority relation. Falsifiable: if real moral reasoning requires priorities that cannot be fixed in advance (context-generated), the approach fails. Handed to Memo 10.

3.9 Handoff to D1/D2

Concrete computational implications for the compiler architects:

1. INFERENCE No Fact → Norm rule in the kernel. Enforce the is/ought type barrier (§3.5.1). The kernel computes prediction (Fact→Fact) and application (Norm ∧ Fact → Norm), never justification.
2. INFERENCE No substantive moral predicate as a kernel primitive. "Justice", "human rights", "the good", a specific virtue list — all excluded by the three-leg argument of §3.6. They enter as provenance-tagged inputs.
3. INFERENCE Carry the content-neutral residue only. The admissible kernel structure is the verdict-function shape (subject, features) → deontic verdict with unbound value parameters (§3.6.4), plus the right/duty pairing and a consistency/closure check.
4. INFERENCE Represent metaethical neutrality explicitly. Per §3.4.5, the correct kernel behavior is invariant across metaethical positions: treat moral norms as typed inputs, never as internally-verifiable axioms. Do not let any D1 component presuppose realism, anti-realism, or constructivism.
5. INFERENCE Type evaluative ("thick") predicates as bound parameters, not descriptive facts (§3.5.2, §3.8). Maintain the registry from §3.8.
6. OPEN → design flag Plan for defeasibility and the O/P/F gap. The kernel's modality set may be insufficient for prima facie duties and supererogation (§3.7); coordinate the resolution with Memo 10's vocabulary stress test before freezing D1's modality type.
7. INFERENCE Keep aggregation of conflicting moral sources above the kernel. Any weighting is itself a value choice (§3.7); the kernel must expose a conflict, not silently resolve it. Coordinate with Memo 01's aggregation limits and Memo 12's interface.

One-line summary: ethics contributes a content-neutral verdict-function shape and a hard is/ought type barrier to the kernel; every substantive ethical commitment stays above it as provenance-tagged, mutable input.